Particle Filters for Graphical Models

Abstract

This paper discloses a novel algorithm for efficient inference in undirected graphical models using Sequential Monte Carlo (SMC) based numerical approximation techniques. The developed methodology extends the applicability of the much celebrated Loopy Belief Propagation (LBP) algorithm to nonlinear, non-Gaussian models, whilst retaining a computational cost that is linear in the number of sample points (or particles). The work presented is thus a general framework that can be applied to a plethora of novel non-linear signal processing problems. In this paper, we apply our inference algorithm to the (sequential problem of) articulated object tracking.